ADK Single-turn Agent as Sub-agent Sample

Overview

This sample demonstrates how a "single_turn" mode agent can act as an autonomous sub-agent to an LLM agent, utilizing schemas and tools without ever interacting with the user.

Note: This is the recommended mechanism to replace the older AgentTool pattern. Unlike AgentTool, using a single_turn sub-agent preserves the sub-agent's internal interactions (like tool calls) in the session history.

Single-turn agents are designed to execute their function fully in one prompt-response cycle. In this sample:

1. phone_recommender: A single-turn agent that receives structured input (UserPreferences), uses a mocked tool (check_phone_price), and returns structured output (PhoneRecommendation).
2. root_agent: The main agent that interacts with the user, translates their natural language request into the structured UserPreferences, and delegates to phone_recommender.

Sample Inputs

• I need a phone mostly for gaming. I have about $1000 to spend.
• What is a good cheap phone from Google for basic tasks?
• I love photography but prefer smaller phones. My budget is $600.

Graph

               [ root_agent ]
                      |
                      v
             [ phone_recommender ] --(uses)--> [ check_phone_price (tool) ]

How To

1. Define a sub-agent with mode="single_turn", input_schema, output_schema:

   pythonCopy

   phone_recommender = Agent(
       name="phone_recommender",
       mode="single_turn",
       input_schema=UserPreferences,
       output_schema=PhoneRecommendation,
       tools=[check_phone_price],
       ...
   )
   

2. Assign it to a parent agent:

   pythonCopy

   root_agent = Agent(
       sub_agents=[phone_recommender],
       ...
   )